1 //===-- X86AsmBackend.cpp - X86 Assembler Backend -------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 9 #include "MCTargetDesc/X86BaseInfo.h" 10 #include "MCTargetDesc/X86FixupKinds.h" 11 #include "llvm/ADT/StringSwitch.h" 12 #include "llvm/BinaryFormat/ELF.h" 13 #include "llvm/BinaryFormat/MachO.h" 14 #include "llvm/MC/MCAsmBackend.h" 15 #include "llvm/MC/MCAsmLayout.h" 16 #include "llvm/MC/MCAssembler.h" 17 #include "llvm/MC/MCCodeEmitter.h" 18 #include "llvm/MC/MCContext.h" 19 #include "llvm/MC/MCDwarf.h" 20 #include "llvm/MC/MCELFObjectWriter.h" 21 #include "llvm/MC/MCExpr.h" 22 #include "llvm/MC/MCFixupKindInfo.h" 23 #include "llvm/MC/MCInst.h" 24 #include "llvm/MC/MCInstrInfo.h" 25 #include "llvm/MC/MCMachObjectWriter.h" 26 #include "llvm/MC/MCObjectStreamer.h" 27 #include "llvm/MC/MCObjectWriter.h" 28 #include "llvm/MC/MCRegisterInfo.h" 29 #include "llvm/MC/MCSectionMachO.h" 30 #include "llvm/MC/MCSubtargetInfo.h" 31 #include "llvm/MC/MCValue.h" 32 #include "llvm/Support/CommandLine.h" 33 #include "llvm/Support/ErrorHandling.h" 34 #include "llvm/Support/TargetRegistry.h" 35 #include "llvm/Support/raw_ostream.h" 36 37 using namespace llvm; 38 39 namespace { 40 /// A wrapper for holding a mask of the values from X86::AlignBranchBoundaryKind 41 class X86AlignBranchKind { 42 private: 43 uint8_t AlignBranchKind = 0; 44 45 public: 46 void operator=(const std::string &Val) { 47 if (Val.empty()) 48 return; 49 SmallVector<StringRef, 6> BranchTypes; 50 StringRef(Val).split(BranchTypes, '+', -1, false); 51 for (auto BranchType : BranchTypes) { 52 if (BranchType == "fused") 53 addKind(X86::AlignBranchFused); 54 else if (BranchType == "jcc") 55 addKind(X86::AlignBranchJcc); 56 else if (BranchType == "jmp") 57 addKind(X86::AlignBranchJmp); 58 else if (BranchType == "call") 59 addKind(X86::AlignBranchCall); 60 else if (BranchType == "ret") 61 addKind(X86::AlignBranchRet); 62 else if (BranchType == "indirect") 63 addKind(X86::AlignBranchIndirect); 64 else { 65 errs() << "invalid argument " << BranchType.str() 66 << " to -x86-align-branch=; each element must be one of: fused, " 67 "jcc, jmp, call, ret, indirect.(plus separated)\n"; 68 } 69 } 70 } 71 72 operator uint8_t() const { return AlignBranchKind; } 73 void addKind(X86::AlignBranchBoundaryKind Value) { AlignBranchKind |= Value; } 74 }; 75 76 X86AlignBranchKind X86AlignBranchKindLoc; 77 78 cl::opt<unsigned> X86AlignBranchBoundary( 79 "x86-align-branch-boundary", cl::init(0), 80 cl::desc( 81 "Control how the assembler should align branches with NOP. If the " 82 "boundary's size is not 0, it should be a power of 2 and no less " 83 "than 32. Branches will be aligned to prevent from being across or " 84 "against the boundary of specified size. The default value 0 does not " 85 "align branches.")); 86 87 cl::opt<X86AlignBranchKind, true, cl::parser<std::string>> X86AlignBranch( 88 "x86-align-branch", 89 cl::desc( 90 "Specify types of branches to align (plus separated list of types):" 91 "\njcc indicates conditional jumps" 92 "\nfused indicates fused conditional jumps" 93 "\njmp indicates direct unconditional jumps" 94 "\ncall indicates direct and indirect calls" 95 "\nret indicates rets" 96 "\nindirect indicates indirect unconditional jumps"), 97 cl::location(X86AlignBranchKindLoc)); 98 99 cl::opt<bool> X86AlignBranchWithin32BBoundaries( 100 "x86-branches-within-32B-boundaries", cl::init(false), 101 cl::desc( 102 "Align selected instructions to mitigate negative performance impact " 103 "of Intel's micro code update for errata skx102. May break " 104 "assumptions about labels corresponding to particular instructions, " 105 "and should be used with caution.")); 106 107 cl::opt<unsigned> X86PadMaxPrefixSize( 108 "x86-pad-max-prefix-size", cl::init(0), 109 cl::desc("Maximum number of prefixes to use for padding")); 110 111 cl::opt<bool> X86PadForAlign( 112 "x86-pad-for-align", cl::init(false), cl::Hidden, 113 cl::desc("Pad previous instructions to implement align directives")); 114 115 cl::opt<bool> X86PadForBranchAlign( 116 "x86-pad-for-branch-align", cl::init(true), cl::Hidden, 117 cl::desc("Pad previous instructions to implement branch alignment")); 118 119 class X86ELFObjectWriter : public MCELFObjectTargetWriter { 120 public: 121 X86ELFObjectWriter(bool is64Bit, uint8_t OSABI, uint16_t EMachine, 122 bool HasRelocationAddend, bool foobar) 123 : MCELFObjectTargetWriter(is64Bit, OSABI, EMachine, HasRelocationAddend) {} 124 }; 125 126 class X86AsmBackend : public MCAsmBackend { 127 const MCSubtargetInfo &STI; 128 std::unique_ptr<const MCInstrInfo> MCII; 129 X86AlignBranchKind AlignBranchType; 130 Align AlignBoundary; 131 unsigned TargetPrefixMax = 0; 132 133 MCInst PrevInst; 134 MCBoundaryAlignFragment *PendingBA = nullptr; 135 std::pair<MCFragment *, size_t> PrevInstPosition; 136 bool CanPadInst; 137 138 uint8_t determinePaddingPrefix(const MCInst &Inst) const; 139 bool isMacroFused(const MCInst &Cmp, const MCInst &Jcc) const; 140 bool needAlign(const MCInst &Inst) const; 141 bool canPadBranches(MCObjectStreamer &OS) const; 142 bool canPadInst(const MCInst &Inst, MCObjectStreamer &OS) const; 143 144 public: 145 X86AsmBackend(const Target &T, const MCSubtargetInfo &STI) 146 : MCAsmBackend(support::little), STI(STI), 147 MCII(T.createMCInstrInfo()) { 148 if (X86AlignBranchWithin32BBoundaries) { 149 // At the moment, this defaults to aligning fused branches, unconditional 150 // jumps, and (unfused) conditional jumps with nops. Both the 151 // instructions aligned and the alignment method (nop vs prefix) may 152 // change in the future. 153 AlignBoundary = assumeAligned(32);; 154 AlignBranchType.addKind(X86::AlignBranchFused); 155 AlignBranchType.addKind(X86::AlignBranchJcc); 156 AlignBranchType.addKind(X86::AlignBranchJmp); 157 } 158 // Allow overriding defaults set by master flag 159 if (X86AlignBranchBoundary.getNumOccurrences()) 160 AlignBoundary = assumeAligned(X86AlignBranchBoundary); 161 if (X86AlignBranch.getNumOccurrences()) 162 AlignBranchType = X86AlignBranchKindLoc; 163 if (X86PadMaxPrefixSize.getNumOccurrences()) 164 TargetPrefixMax = X86PadMaxPrefixSize; 165 } 166 167 bool allowAutoPadding() const override; 168 bool allowEnhancedRelaxation() const override; 169 void emitInstructionBegin(MCObjectStreamer &OS, const MCInst &Inst) override; 170 void emitInstructionEnd(MCObjectStreamer &OS, const MCInst &Inst) override; 171 172 unsigned getNumFixupKinds() const override { 173 return X86::NumTargetFixupKinds; 174 } 175 176 Optional<MCFixupKind> getFixupKind(StringRef Name) const override; 177 178 const MCFixupKindInfo &getFixupKindInfo(MCFixupKind Kind) const override; 179 180 bool shouldForceRelocation(const MCAssembler &Asm, const MCFixup &Fixup, 181 const MCValue &Target) override; 182 183 void applyFixup(const MCAssembler &Asm, const MCFixup &Fixup, 184 const MCValue &Target, MutableArrayRef<char> Data, 185 uint64_t Value, bool IsResolved, 186 const MCSubtargetInfo *STI) const override; 187 188 bool mayNeedRelaxation(const MCInst &Inst, 189 const MCSubtargetInfo &STI) const override; 190 191 bool fixupNeedsRelaxation(const MCFixup &Fixup, uint64_t Value, 192 const MCRelaxableFragment *DF, 193 const MCAsmLayout &Layout) const override; 194 195 void relaxInstruction(MCInst &Inst, 196 const MCSubtargetInfo &STI) const override; 197 198 bool padInstructionViaRelaxation(MCRelaxableFragment &RF, 199 MCCodeEmitter &Emitter, 200 unsigned &RemainingSize) const; 201 202 bool padInstructionViaPrefix(MCRelaxableFragment &RF, MCCodeEmitter &Emitter, 203 unsigned &RemainingSize) const; 204 205 bool padInstructionEncoding(MCRelaxableFragment &RF, MCCodeEmitter &Emitter, 206 unsigned &RemainingSize) const; 207 208 void finishLayout(MCAssembler const &Asm, MCAsmLayout &Layout) const override; 209 210 unsigned getMaximumNopSize() const override; 211 212 bool writeNopData(raw_ostream &OS, uint64_t Count) const override; 213 }; 214 } // end anonymous namespace 215 216 static unsigned getRelaxedOpcodeBranch(const MCInst &Inst, bool Is16BitMode) { 217 unsigned Op = Inst.getOpcode(); 218 switch (Op) { 219 default: 220 return Op; 221 case X86::JCC_1: 222 return (Is16BitMode) ? X86::JCC_2 : X86::JCC_4; 223 case X86::JMP_1: 224 return (Is16BitMode) ? X86::JMP_2 : X86::JMP_4; 225 } 226 } 227 228 static unsigned getRelaxedOpcodeArith(const MCInst &Inst) { 229 unsigned Op = Inst.getOpcode(); 230 switch (Op) { 231 default: 232 return Op; 233 234 // IMUL 235 case X86::IMUL16rri8: return X86::IMUL16rri; 236 case X86::IMUL16rmi8: return X86::IMUL16rmi; 237 case X86::IMUL32rri8: return X86::IMUL32rri; 238 case X86::IMUL32rmi8: return X86::IMUL32rmi; 239 case X86::IMUL64rri8: return X86::IMUL64rri32; 240 case X86::IMUL64rmi8: return X86::IMUL64rmi32; 241 242 // AND 243 case X86::AND16ri8: return X86::AND16ri; 244 case X86::AND16mi8: return X86::AND16mi; 245 case X86::AND32ri8: return X86::AND32ri; 246 case X86::AND32mi8: return X86::AND32mi; 247 case X86::AND64ri8: return X86::AND64ri32; 248 case X86::AND64mi8: return X86::AND64mi32; 249 250 // OR 251 case X86::OR16ri8: return X86::OR16ri; 252 case X86::OR16mi8: return X86::OR16mi; 253 case X86::OR32ri8: return X86::OR32ri; 254 case X86::OR32mi8: return X86::OR32mi; 255 case X86::OR64ri8: return X86::OR64ri32; 256 case X86::OR64mi8: return X86::OR64mi32; 257 258 // XOR 259 case X86::XOR16ri8: return X86::XOR16ri; 260 case X86::XOR16mi8: return X86::XOR16mi; 261 case X86::XOR32ri8: return X86::XOR32ri; 262 case X86::XOR32mi8: return X86::XOR32mi; 263 case X86::XOR64ri8: return X86::XOR64ri32; 264 case X86::XOR64mi8: return X86::XOR64mi32; 265 266 // ADD 267 case X86::ADD16ri8: return X86::ADD16ri; 268 case X86::ADD16mi8: return X86::ADD16mi; 269 case X86::ADD32ri8: return X86::ADD32ri; 270 case X86::ADD32mi8: return X86::ADD32mi; 271 case X86::ADD64ri8: return X86::ADD64ri32; 272 case X86::ADD64mi8: return X86::ADD64mi32; 273 274 // ADC 275 case X86::ADC16ri8: return X86::ADC16ri; 276 case X86::ADC16mi8: return X86::ADC16mi; 277 case X86::ADC32ri8: return X86::ADC32ri; 278 case X86::ADC32mi8: return X86::ADC32mi; 279 case X86::ADC64ri8: return X86::ADC64ri32; 280 case X86::ADC64mi8: return X86::ADC64mi32; 281 282 // SUB 283 case X86::SUB16ri8: return X86::SUB16ri; 284 case X86::SUB16mi8: return X86::SUB16mi; 285 case X86::SUB32ri8: return X86::SUB32ri; 286 case X86::SUB32mi8: return X86::SUB32mi; 287 case X86::SUB64ri8: return X86::SUB64ri32; 288 case X86::SUB64mi8: return X86::SUB64mi32; 289 290 // SBB 291 case X86::SBB16ri8: return X86::SBB16ri; 292 case X86::SBB16mi8: return X86::SBB16mi; 293 case X86::SBB32ri8: return X86::SBB32ri; 294 case X86::SBB32mi8: return X86::SBB32mi; 295 case X86::SBB64ri8: return X86::SBB64ri32; 296 case X86::SBB64mi8: return X86::SBB64mi32; 297 298 // CMP 299 case X86::CMP16ri8: return X86::CMP16ri; 300 case X86::CMP16mi8: return X86::CMP16mi; 301 case X86::CMP32ri8: return X86::CMP32ri; 302 case X86::CMP32mi8: return X86::CMP32mi; 303 case X86::CMP64ri8: return X86::CMP64ri32; 304 case X86::CMP64mi8: return X86::CMP64mi32; 305 306 // PUSH 307 case X86::PUSH32i8: return X86::PUSHi32; 308 case X86::PUSH16i8: return X86::PUSHi16; 309 case X86::PUSH64i8: return X86::PUSH64i32; 310 } 311 } 312 313 static unsigned getRelaxedOpcode(const MCInst &Inst, bool Is16BitMode) { 314 unsigned R = getRelaxedOpcodeArith(Inst); 315 if (R != Inst.getOpcode()) 316 return R; 317 return getRelaxedOpcodeBranch(Inst, Is16BitMode); 318 } 319 320 static X86::CondCode getCondFromBranch(const MCInst &MI, 321 const MCInstrInfo &MCII) { 322 unsigned Opcode = MI.getOpcode(); 323 switch (Opcode) { 324 default: 325 return X86::COND_INVALID; 326 case X86::JCC_1: { 327 const MCInstrDesc &Desc = MCII.get(Opcode); 328 return static_cast<X86::CondCode>( 329 MI.getOperand(Desc.getNumOperands() - 1).getImm()); 330 } 331 } 332 } 333 334 static X86::SecondMacroFusionInstKind 335 classifySecondInstInMacroFusion(const MCInst &MI, const MCInstrInfo &MCII) { 336 X86::CondCode CC = getCondFromBranch(MI, MCII); 337 return classifySecondCondCodeInMacroFusion(CC); 338 } 339 340 /// Check if the instruction uses RIP relative addressing. 341 static bool isRIPRelative(const MCInst &MI, const MCInstrInfo &MCII) { 342 unsigned Opcode = MI.getOpcode(); 343 const MCInstrDesc &Desc = MCII.get(Opcode); 344 uint64_t TSFlags = Desc.TSFlags; 345 unsigned CurOp = X86II::getOperandBias(Desc); 346 int MemoryOperand = X86II::getMemoryOperandNo(TSFlags); 347 if (MemoryOperand < 0) 348 return false; 349 unsigned BaseRegNum = MemoryOperand + CurOp + X86::AddrBaseReg; 350 unsigned BaseReg = MI.getOperand(BaseRegNum).getReg(); 351 return (BaseReg == X86::RIP); 352 } 353 354 /// Check if the instruction is a prefix. 355 static bool isPrefix(const MCInst &MI, const MCInstrInfo &MCII) { 356 return X86II::isPrefix(MCII.get(MI.getOpcode()).TSFlags); 357 } 358 359 /// Check if the instruction is valid as the first instruction in macro fusion. 360 static bool isFirstMacroFusibleInst(const MCInst &Inst, 361 const MCInstrInfo &MCII) { 362 // An Intel instruction with RIP relative addressing is not macro fusible. 363 if (isRIPRelative(Inst, MCII)) 364 return false; 365 X86::FirstMacroFusionInstKind FIK = 366 X86::classifyFirstOpcodeInMacroFusion(Inst.getOpcode()); 367 return FIK != X86::FirstMacroFusionInstKind::Invalid; 368 } 369 370 /// X86 can reduce the bytes of NOP by padding instructions with prefixes to 371 /// get a better peformance in some cases. Here, we determine which prefix is 372 /// the most suitable. 373 /// 374 /// If the instruction has a segment override prefix, use the existing one. 375 /// If the target is 64-bit, use the CS. 376 /// If the target is 32-bit, 377 /// - If the instruction has a ESP/EBP base register, use SS. 378 /// - Otherwise use DS. 379 uint8_t X86AsmBackend::determinePaddingPrefix(const MCInst &Inst) const { 380 assert((STI.hasFeature(X86::Mode32Bit) || STI.hasFeature(X86::Mode64Bit)) && 381 "Prefixes can be added only in 32-bit or 64-bit mode."); 382 const MCInstrDesc &Desc = MCII->get(Inst.getOpcode()); 383 uint64_t TSFlags = Desc.TSFlags; 384 385 // Determine where the memory operand starts, if present. 386 int MemoryOperand = X86II::getMemoryOperandNo(TSFlags); 387 if (MemoryOperand != -1) 388 MemoryOperand += X86II::getOperandBias(Desc); 389 390 unsigned SegmentReg = 0; 391 if (MemoryOperand >= 0) { 392 // Check for explicit segment override on memory operand. 393 SegmentReg = Inst.getOperand(MemoryOperand + X86::AddrSegmentReg).getReg(); 394 } 395 396 switch (TSFlags & X86II::FormMask) { 397 default: 398 break; 399 case X86II::RawFrmDstSrc: { 400 // Check segment override opcode prefix as needed (not for %ds). 401 if (Inst.getOperand(2).getReg() != X86::DS) 402 SegmentReg = Inst.getOperand(2).getReg(); 403 break; 404 } 405 case X86II::RawFrmSrc: { 406 // Check segment override opcode prefix as needed (not for %ds). 407 if (Inst.getOperand(1).getReg() != X86::DS) 408 SegmentReg = Inst.getOperand(1).getReg(); 409 break; 410 } 411 case X86II::RawFrmMemOffs: { 412 // Check segment override opcode prefix as needed. 413 SegmentReg = Inst.getOperand(1).getReg(); 414 break; 415 } 416 } 417 418 if (SegmentReg != 0) 419 return X86::getSegmentOverridePrefixForReg(SegmentReg); 420 421 if (STI.hasFeature(X86::Mode64Bit)) 422 return X86::CS_Encoding; 423 424 if (MemoryOperand >= 0) { 425 unsigned BaseRegNum = MemoryOperand + X86::AddrBaseReg; 426 unsigned BaseReg = Inst.getOperand(BaseRegNum).getReg(); 427 if (BaseReg == X86::ESP || BaseReg == X86::EBP) 428 return X86::SS_Encoding; 429 } 430 return X86::DS_Encoding; 431 } 432 433 /// Check if the two instructions will be macro-fused on the target cpu. 434 bool X86AsmBackend::isMacroFused(const MCInst &Cmp, const MCInst &Jcc) const { 435 const MCInstrDesc &InstDesc = MCII->get(Jcc.getOpcode()); 436 if (!InstDesc.isConditionalBranch()) 437 return false; 438 if (!isFirstMacroFusibleInst(Cmp, *MCII)) 439 return false; 440 const X86::FirstMacroFusionInstKind CmpKind = 441 X86::classifyFirstOpcodeInMacroFusion(Cmp.getOpcode()); 442 const X86::SecondMacroFusionInstKind BranchKind = 443 classifySecondInstInMacroFusion(Jcc, *MCII); 444 return X86::isMacroFused(CmpKind, BranchKind); 445 } 446 447 /// Check if the instruction has a variant symbol operand. 448 static bool hasVariantSymbol(const MCInst &MI) { 449 for (auto &Operand : MI) { 450 if (!Operand.isExpr()) 451 continue; 452 const MCExpr &Expr = *Operand.getExpr(); 453 if (Expr.getKind() == MCExpr::SymbolRef && 454 cast<MCSymbolRefExpr>(Expr).getKind() != MCSymbolRefExpr::VK_None) 455 return true; 456 } 457 return false; 458 } 459 460 bool X86AsmBackend::allowAutoPadding() const { 461 return (AlignBoundary != Align(1) && AlignBranchType != X86::AlignBranchNone); 462 } 463 464 bool X86AsmBackend::allowEnhancedRelaxation() const { 465 return allowAutoPadding() && TargetPrefixMax != 0 && X86PadForBranchAlign; 466 } 467 468 /// X86 has certain instructions which enable interrupts exactly one 469 /// instruction *after* the instruction which stores to SS. Return true if the 470 /// given instruction has such an interrupt delay slot. 471 static bool hasInterruptDelaySlot(const MCInst &Inst) { 472 switch (Inst.getOpcode()) { 473 case X86::POPSS16: 474 case X86::POPSS32: 475 case X86::STI: 476 return true; 477 478 case X86::MOV16sr: 479 case X86::MOV32sr: 480 case X86::MOV64sr: 481 case X86::MOV16sm: 482 if (Inst.getOperand(0).getReg() == X86::SS) 483 return true; 484 break; 485 } 486 return false; 487 } 488 489 /// Check if the instruction to be emitted is right after any data. 490 static bool 491 isRightAfterData(MCFragment *CurrentFragment, 492 const std::pair<MCFragment *, size_t> &PrevInstPosition) { 493 MCFragment *F = CurrentFragment; 494 // Empty data fragments may be created to prevent further data being 495 // added into the previous fragment, we need to skip them since they 496 // have no contents. 497 for (; isa_and_nonnull<MCDataFragment>(F); F = F->getPrevNode()) 498 if (cast<MCDataFragment>(F)->getContents().size() != 0) 499 break; 500 501 // Since data is always emitted into a DataFragment, our check strategy is 502 // simple here. 503 // - If the fragment is a DataFragment 504 // - If it's not the fragment where the previous instruction is, 505 // returns true. 506 // - If it's the fragment holding the previous instruction but its 507 // size changed since the the previous instruction was emitted into 508 // it, returns true. 509 // - Otherwise returns false. 510 // - If the fragment is not a DataFragment, returns false. 511 if (auto *DF = dyn_cast_or_null<MCDataFragment>(F)) 512 return DF != PrevInstPosition.first || 513 DF->getContents().size() != PrevInstPosition.second; 514 515 return false; 516 } 517 518 /// \returns the fragment size if it has instructions, otherwise returns 0. 519 static size_t getSizeForInstFragment(const MCFragment *F) { 520 if (!F || !F->hasInstructions()) 521 return 0; 522 // MCEncodedFragmentWithContents being templated makes this tricky. 523 switch (F->getKind()) { 524 default: 525 llvm_unreachable("Unknown fragment with instructions!"); 526 case MCFragment::FT_Data: 527 return cast<MCDataFragment>(*F).getContents().size(); 528 case MCFragment::FT_Relaxable: 529 return cast<MCRelaxableFragment>(*F).getContents().size(); 530 case MCFragment::FT_CompactEncodedInst: 531 return cast<MCCompactEncodedInstFragment>(*F).getContents().size(); 532 } 533 } 534 535 /// Return true if we can insert NOP or prefixes automatically before the 536 /// the instruction to be emitted. 537 bool X86AsmBackend::canPadInst(const MCInst &Inst, MCObjectStreamer &OS) const { 538 if (hasVariantSymbol(Inst)) 539 // Linker may rewrite the instruction with variant symbol operand(e.g. 540 // TLSCALL). 541 return false; 542 543 if (hasInterruptDelaySlot(PrevInst)) 544 // If this instruction follows an interrupt enabling instruction with a one 545 // instruction delay, inserting a nop would change behavior. 546 return false; 547 548 if (isPrefix(PrevInst, *MCII)) 549 // If this instruction follows a prefix, inserting a nop/prefix would change 550 // semantic. 551 return false; 552 553 if (isPrefix(Inst, *MCII)) 554 // If this instruction is a prefix, inserting a prefix would change 555 // semantic. 556 return false; 557 558 if (isRightAfterData(OS.getCurrentFragment(), PrevInstPosition)) 559 // If this instruction follows any data, there is no clear 560 // instruction boundary, inserting a nop/prefix would change semantic. 561 return false; 562 563 return true; 564 } 565 566 bool X86AsmBackend::canPadBranches(MCObjectStreamer &OS) const { 567 if (!OS.getAllowAutoPadding()) 568 return false; 569 assert(allowAutoPadding() && "incorrect initialization!"); 570 571 // We only pad in text section. 572 if (!OS.getCurrentSectionOnly()->getKind().isText()) 573 return false; 574 575 // To be Done: Currently don't deal with Bundle cases. 576 if (OS.getAssembler().isBundlingEnabled()) 577 return false; 578 579 // Branches only need to be aligned in 32-bit or 64-bit mode. 580 if (!(STI.hasFeature(X86::Mode64Bit) || STI.hasFeature(X86::Mode32Bit))) 581 return false; 582 583 return true; 584 } 585 586 /// Check if the instruction operand needs to be aligned. 587 bool X86AsmBackend::needAlign(const MCInst &Inst) const { 588 const MCInstrDesc &Desc = MCII->get(Inst.getOpcode()); 589 return (Desc.isConditionalBranch() && 590 (AlignBranchType & X86::AlignBranchJcc)) || 591 (Desc.isUnconditionalBranch() && 592 (AlignBranchType & X86::AlignBranchJmp)) || 593 (Desc.isCall() && (AlignBranchType & X86::AlignBranchCall)) || 594 (Desc.isReturn() && (AlignBranchType & X86::AlignBranchRet)) || 595 (Desc.isIndirectBranch() && 596 (AlignBranchType & X86::AlignBranchIndirect)); 597 } 598 599 /// Insert BoundaryAlignFragment before instructions to align branches. 600 void X86AsmBackend::emitInstructionBegin(MCObjectStreamer &OS, 601 const MCInst &Inst) { 602 CanPadInst = canPadInst(Inst, OS); 603 604 if (!canPadBranches(OS)) 605 return; 606 607 if (!isMacroFused(PrevInst, Inst)) 608 // Macro fusion doesn't happen indeed, clear the pending. 609 PendingBA = nullptr; 610 611 if (!CanPadInst) 612 return; 613 614 if (PendingBA && OS.getCurrentFragment()->getPrevNode() == PendingBA) { 615 // Macro fusion actually happens and there is no other fragment inserted 616 // after the previous instruction. 617 // 618 // Do nothing here since we already inserted a BoudaryAlign fragment when 619 // we met the first instruction in the fused pair and we'll tie them 620 // together in emitInstructionEnd. 621 // 622 // Note: When there is at least one fragment, such as MCAlignFragment, 623 // inserted after the previous instruction, e.g. 624 // 625 // \code 626 // cmp %rax %rcx 627 // .align 16 628 // je .Label0 629 // \ endcode 630 // 631 // We will treat the JCC as a unfused branch although it may be fused 632 // with the CMP. 633 return; 634 } 635 636 if (needAlign(Inst) || ((AlignBranchType & X86::AlignBranchFused) && 637 isFirstMacroFusibleInst(Inst, *MCII))) { 638 // If we meet a unfused branch or the first instuction in a fusiable pair, 639 // insert a BoundaryAlign fragment. 640 OS.insert(PendingBA = new MCBoundaryAlignFragment(AlignBoundary)); 641 } 642 } 643 644 /// Set the last fragment to be aligned for the BoundaryAlignFragment. 645 void X86AsmBackend::emitInstructionEnd(MCObjectStreamer &OS, const MCInst &Inst) { 646 PrevInst = Inst; 647 MCFragment *CF = OS.getCurrentFragment(); 648 PrevInstPosition = std::make_pair(CF, getSizeForInstFragment(CF)); 649 if (auto *F = dyn_cast_or_null<MCRelaxableFragment>(CF)) 650 F->setAllowAutoPadding(CanPadInst); 651 652 if (!canPadBranches(OS)) 653 return; 654 655 if (!needAlign(Inst) || !PendingBA) 656 return; 657 658 // Tie the aligned instructions into a a pending BoundaryAlign. 659 PendingBA->setLastFragment(CF); 660 PendingBA = nullptr; 661 662 // We need to ensure that further data isn't added to the current 663 // DataFragment, so that we can get the size of instructions later in 664 // MCAssembler::relaxBoundaryAlign. The easiest way is to insert a new empty 665 // DataFragment. 666 if (isa_and_nonnull<MCDataFragment>(CF)) 667 OS.insert(new MCDataFragment()); 668 669 // Update the maximum alignment on the current section if necessary. 670 MCSection *Sec = OS.getCurrentSectionOnly(); 671 if (AlignBoundary.value() > Sec->getAlignment()) 672 Sec->setAlignment(AlignBoundary); 673 } 674 675 Optional<MCFixupKind> X86AsmBackend::getFixupKind(StringRef Name) const { 676 if (STI.getTargetTriple().isOSBinFormatELF()) { 677 unsigned Type; 678 if (STI.getTargetTriple().getArch() == Triple::x86_64) { 679 Type = llvm::StringSwitch<unsigned>(Name) 680 #define ELF_RELOC(X, Y) .Case(#X, Y) 681 #include "llvm/BinaryFormat/ELFRelocs/x86_64.def" 682 #undef ELF_RELOC 683 .Case("BFD_RELOC_NONE", ELF::R_X86_64_NONE) 684 .Case("BFD_RELOC_8", ELF::R_X86_64_8) 685 .Case("BFD_RELOC_16", ELF::R_X86_64_16) 686 .Case("BFD_RELOC_32", ELF::R_X86_64_32) 687 .Case("BFD_RELOC_64", ELF::R_X86_64_64) 688 .Default(-1u); 689 } else { 690 Type = llvm::StringSwitch<unsigned>(Name) 691 #define ELF_RELOC(X, Y) .Case(#X, Y) 692 #include "llvm/BinaryFormat/ELFRelocs/i386.def" 693 #undef ELF_RELOC 694 .Case("BFD_RELOC_NONE", ELF::R_386_NONE) 695 .Case("BFD_RELOC_8", ELF::R_386_8) 696 .Case("BFD_RELOC_16", ELF::R_386_16) 697 .Case("BFD_RELOC_32", ELF::R_386_32) 698 .Default(-1u); 699 } 700 if (Type == -1u) 701 return None; 702 return static_cast<MCFixupKind>(FirstLiteralRelocationKind + Type); 703 } 704 return MCAsmBackend::getFixupKind(Name); 705 } 706 707 const MCFixupKindInfo &X86AsmBackend::getFixupKindInfo(MCFixupKind Kind) const { 708 const static MCFixupKindInfo Infos[X86::NumTargetFixupKinds] = { 709 {"reloc_riprel_4byte", 0, 32, MCFixupKindInfo::FKF_IsPCRel}, 710 {"reloc_riprel_4byte_movq_load", 0, 32, MCFixupKindInfo::FKF_IsPCRel}, 711 {"reloc_riprel_4byte_relax", 0, 32, MCFixupKindInfo::FKF_IsPCRel}, 712 {"reloc_riprel_4byte_relax_rex", 0, 32, MCFixupKindInfo::FKF_IsPCRel}, 713 {"reloc_signed_4byte", 0, 32, 0}, 714 {"reloc_signed_4byte_relax", 0, 32, 0}, 715 {"reloc_global_offset_table", 0, 32, 0}, 716 {"reloc_global_offset_table8", 0, 64, 0}, 717 {"reloc_branch_4byte_pcrel", 0, 32, MCFixupKindInfo::FKF_IsPCRel}, 718 }; 719 720 // Fixup kinds from .reloc directive are like R_386_NONE/R_X86_64_NONE. They 721 // do not require any extra processing. 722 if (Kind >= FirstLiteralRelocationKind) 723 return MCAsmBackend::getFixupKindInfo(FK_NONE); 724 725 if (Kind < FirstTargetFixupKind) 726 return MCAsmBackend::getFixupKindInfo(Kind); 727 728 assert(unsigned(Kind - FirstTargetFixupKind) < getNumFixupKinds() && 729 "Invalid kind!"); 730 assert(Infos[Kind - FirstTargetFixupKind].Name && "Empty fixup name!"); 731 return Infos[Kind - FirstTargetFixupKind]; 732 } 733 734 bool X86AsmBackend::shouldForceRelocation(const MCAssembler &, 735 const MCFixup &Fixup, 736 const MCValue &) { 737 return Fixup.getKind() >= FirstLiteralRelocationKind; 738 } 739 740 static unsigned getFixupKindSize(unsigned Kind) { 741 switch (Kind) { 742 default: 743 llvm_unreachable("invalid fixup kind!"); 744 case FK_NONE: 745 return 0; 746 case FK_PCRel_1: 747 case FK_SecRel_1: 748 case FK_Data_1: 749 return 1; 750 case FK_PCRel_2: 751 case FK_SecRel_2: 752 case FK_Data_2: 753 return 2; 754 case FK_PCRel_4: 755 case X86::reloc_riprel_4byte: 756 case X86::reloc_riprel_4byte_relax: 757 case X86::reloc_riprel_4byte_relax_rex: 758 case X86::reloc_riprel_4byte_movq_load: 759 case X86::reloc_signed_4byte: 760 case X86::reloc_signed_4byte_relax: 761 case X86::reloc_global_offset_table: 762 case X86::reloc_branch_4byte_pcrel: 763 case FK_SecRel_4: 764 case FK_Data_4: 765 return 4; 766 case FK_PCRel_8: 767 case FK_SecRel_8: 768 case FK_Data_8: 769 case X86::reloc_global_offset_table8: 770 return 8; 771 } 772 } 773 774 void X86AsmBackend::applyFixup(const MCAssembler &Asm, const MCFixup &Fixup, 775 const MCValue &Target, 776 MutableArrayRef<char> Data, 777 uint64_t Value, bool IsResolved, 778 const MCSubtargetInfo *STI) const { 779 unsigned Kind = Fixup.getKind(); 780 if (Kind >= FirstLiteralRelocationKind) 781 return; 782 unsigned Size = getFixupKindSize(Kind); 783 784 assert(Fixup.getOffset() + Size <= Data.size() && "Invalid fixup offset!"); 785 786 int64_t SignedValue = static_cast<int64_t>(Value); 787 if ((Target.isAbsolute() || IsResolved) && 788 getFixupKindInfo(Fixup.getKind()).Flags & 789 MCFixupKindInfo::FKF_IsPCRel) { 790 // check that PC relative fixup fits into the fixup size. 791 if (Size > 0 && !isIntN(Size * 8, SignedValue)) 792 Asm.getContext().reportError( 793 Fixup.getLoc(), "value of " + Twine(SignedValue) + 794 " is too large for field of " + Twine(Size) + 795 ((Size == 1) ? " byte." : " bytes.")); 796 } else { 797 // Check that uppper bits are either all zeros or all ones. 798 // Specifically ignore overflow/underflow as long as the leakage is 799 // limited to the lower bits. This is to remain compatible with 800 // other assemblers. 801 assert((Size == 0 || isIntN(Size * 8 + 1, SignedValue)) && 802 "Value does not fit in the Fixup field"); 803 } 804 805 for (unsigned i = 0; i != Size; ++i) 806 Data[Fixup.getOffset() + i] = uint8_t(Value >> (i * 8)); 807 } 808 809 bool X86AsmBackend::mayNeedRelaxation(const MCInst &Inst, 810 const MCSubtargetInfo &STI) const { 811 // Branches can always be relaxed in either mode. 812 if (getRelaxedOpcodeBranch(Inst, false) != Inst.getOpcode()) 813 return true; 814 815 // Check if this instruction is ever relaxable. 816 if (getRelaxedOpcodeArith(Inst) == Inst.getOpcode()) 817 return false; 818 819 820 // Check if the relaxable operand has an expression. For the current set of 821 // relaxable instructions, the relaxable operand is always the last operand. 822 unsigned RelaxableOp = Inst.getNumOperands() - 1; 823 if (Inst.getOperand(RelaxableOp).isExpr()) 824 return true; 825 826 return false; 827 } 828 829 bool X86AsmBackend::fixupNeedsRelaxation(const MCFixup &Fixup, 830 uint64_t Value, 831 const MCRelaxableFragment *DF, 832 const MCAsmLayout &Layout) const { 833 // Relax if the value is too big for a (signed) i8. 834 return !isInt<8>(Value); 835 } 836 837 // FIXME: Can tblgen help at all here to verify there aren't other instructions 838 // we can relax? 839 void X86AsmBackend::relaxInstruction(MCInst &Inst, 840 const MCSubtargetInfo &STI) const { 841 // The only relaxations X86 does is from a 1byte pcrel to a 4byte pcrel. 842 bool Is16BitMode = STI.getFeatureBits()[X86::Mode16Bit]; 843 unsigned RelaxedOp = getRelaxedOpcode(Inst, Is16BitMode); 844 845 if (RelaxedOp == Inst.getOpcode()) { 846 SmallString<256> Tmp; 847 raw_svector_ostream OS(Tmp); 848 Inst.dump_pretty(OS); 849 OS << "\n"; 850 report_fatal_error("unexpected instruction to relax: " + OS.str()); 851 } 852 853 Inst.setOpcode(RelaxedOp); 854 } 855 856 /// Return true if this instruction has been fully relaxed into it's most 857 /// general available form. 858 static bool isFullyRelaxed(const MCRelaxableFragment &RF) { 859 auto &Inst = RF.getInst(); 860 auto &STI = *RF.getSubtargetInfo(); 861 bool Is16BitMode = STI.getFeatureBits()[X86::Mode16Bit]; 862 return getRelaxedOpcode(Inst, Is16BitMode) == Inst.getOpcode(); 863 } 864 865 bool X86AsmBackend::padInstructionViaPrefix(MCRelaxableFragment &RF, 866 MCCodeEmitter &Emitter, 867 unsigned &RemainingSize) const { 868 if (!RF.getAllowAutoPadding()) 869 return false; 870 // If the instruction isn't fully relaxed, shifting it around might require a 871 // larger value for one of the fixups then can be encoded. The outer loop 872 // will also catch this before moving to the next instruction, but we need to 873 // prevent padding this single instruction as well. 874 if (!isFullyRelaxed(RF)) 875 return false; 876 877 const unsigned OldSize = RF.getContents().size(); 878 if (OldSize == 15) 879 return false; 880 881 const unsigned MaxPossiblePad = std::min(15 - OldSize, RemainingSize); 882 const unsigned RemainingPrefixSize = [&]() -> unsigned { 883 SmallString<15> Code; 884 raw_svector_ostream VecOS(Code); 885 Emitter.emitPrefix(RF.getInst(), VecOS, STI); 886 assert(Code.size() < 15 && "The number of prefixes must be less than 15."); 887 888 // TODO: It turns out we need a decent amount of plumbing for the target 889 // specific bits to determine number of prefixes its safe to add. Various 890 // targets (older chips mostly, but also Atom family) encounter decoder 891 // stalls with too many prefixes. For testing purposes, we set the value 892 // externally for the moment. 893 unsigned ExistingPrefixSize = Code.size(); 894 if (TargetPrefixMax <= ExistingPrefixSize) 895 return 0; 896 return TargetPrefixMax - ExistingPrefixSize; 897 }(); 898 const unsigned PrefixBytesToAdd = 899 std::min(MaxPossiblePad, RemainingPrefixSize); 900 if (PrefixBytesToAdd == 0) 901 return false; 902 903 const uint8_t Prefix = determinePaddingPrefix(RF.getInst()); 904 905 SmallString<256> Code; 906 Code.append(PrefixBytesToAdd, Prefix); 907 Code.append(RF.getContents().begin(), RF.getContents().end()); 908 RF.getContents() = Code; 909 910 // Adjust the fixups for the change in offsets 911 for (auto &F : RF.getFixups()) { 912 F.setOffset(F.getOffset() + PrefixBytesToAdd); 913 } 914 915 RemainingSize -= PrefixBytesToAdd; 916 return true; 917 } 918 919 bool X86AsmBackend::padInstructionViaRelaxation(MCRelaxableFragment &RF, 920 MCCodeEmitter &Emitter, 921 unsigned &RemainingSize) const { 922 if (isFullyRelaxed(RF)) 923 // TODO: There are lots of other tricks we could apply for increasing 924 // encoding size without impacting performance. 925 return false; 926 927 MCInst Relaxed = RF.getInst(); 928 relaxInstruction(Relaxed, *RF.getSubtargetInfo()); 929 930 SmallVector<MCFixup, 4> Fixups; 931 SmallString<15> Code; 932 raw_svector_ostream VecOS(Code); 933 Emitter.encodeInstruction(Relaxed, VecOS, Fixups, *RF.getSubtargetInfo()); 934 const unsigned OldSize = RF.getContents().size(); 935 const unsigned NewSize = Code.size(); 936 assert(NewSize >= OldSize && "size decrease during relaxation?"); 937 unsigned Delta = NewSize - OldSize; 938 if (Delta > RemainingSize) 939 return false; 940 RF.setInst(Relaxed); 941 RF.getContents() = Code; 942 RF.getFixups() = Fixups; 943 RemainingSize -= Delta; 944 return true; 945 } 946 947 bool X86AsmBackend::padInstructionEncoding(MCRelaxableFragment &RF, 948 MCCodeEmitter &Emitter, 949 unsigned &RemainingSize) const { 950 bool Changed = false; 951 if (RemainingSize != 0) 952 Changed |= padInstructionViaRelaxation(RF, Emitter, RemainingSize); 953 if (RemainingSize != 0) 954 Changed |= padInstructionViaPrefix(RF, Emitter, RemainingSize); 955 return Changed; 956 } 957 958 void X86AsmBackend::finishLayout(MCAssembler const &Asm, 959 MCAsmLayout &Layout) const { 960 // See if we can further relax some instructions to cut down on the number of 961 // nop bytes required for code alignment. The actual win is in reducing 962 // instruction count, not number of bytes. Modern X86-64 can easily end up 963 // decode limited. It is often better to reduce the number of instructions 964 // (i.e. eliminate nops) even at the cost of increasing the size and 965 // complexity of others. 966 if (!X86PadForAlign && !X86PadForBranchAlign) 967 return; 968 969 // The processed regions are delimitered by LabeledFragments. -g may have more 970 // MCSymbols and therefore different relaxation results. X86PadForAlign is 971 // disabled by default to eliminate the -g vs non -g difference. 972 DenseSet<MCFragment *> LabeledFragments; 973 for (const MCSymbol &S : Asm.symbols()) 974 LabeledFragments.insert(S.getFragment(false)); 975 976 for (MCSection &Sec : Asm) { 977 if (!Sec.getKind().isText()) 978 continue; 979 980 SmallVector<MCRelaxableFragment *, 4> Relaxable; 981 for (MCSection::iterator I = Sec.begin(), IE = Sec.end(); I != IE; ++I) { 982 MCFragment &F = *I; 983 984 if (LabeledFragments.count(&F)) 985 Relaxable.clear(); 986 987 if (F.getKind() == MCFragment::FT_Data || 988 F.getKind() == MCFragment::FT_CompactEncodedInst) 989 // Skip and ignore 990 continue; 991 992 if (F.getKind() == MCFragment::FT_Relaxable) { 993 auto &RF = cast<MCRelaxableFragment>(*I); 994 Relaxable.push_back(&RF); 995 continue; 996 } 997 998 auto canHandle = [](MCFragment &F) -> bool { 999 switch (F.getKind()) { 1000 default: 1001 return false; 1002 case MCFragment::FT_Align: 1003 return X86PadForAlign; 1004 case MCFragment::FT_BoundaryAlign: 1005 return X86PadForBranchAlign; 1006 } 1007 }; 1008 // For any unhandled kind, assume we can't change layout. 1009 if (!canHandle(F)) { 1010 Relaxable.clear(); 1011 continue; 1012 } 1013 1014 #ifndef NDEBUG 1015 const uint64_t OrigOffset = Layout.getFragmentOffset(&F); 1016 #endif 1017 const uint64_t OrigSize = Asm.computeFragmentSize(Layout, F); 1018 1019 // To keep the effects local, prefer to relax instructions closest to 1020 // the align directive. This is purely about human understandability 1021 // of the resulting code. If we later find a reason to expand 1022 // particular instructions over others, we can adjust. 1023 MCFragment *FirstChangedFragment = nullptr; 1024 unsigned RemainingSize = OrigSize; 1025 while (!Relaxable.empty() && RemainingSize != 0) { 1026 auto &RF = *Relaxable.pop_back_val(); 1027 // Give the backend a chance to play any tricks it wishes to increase 1028 // the encoding size of the given instruction. Target independent code 1029 // will try further relaxation, but target's may play further tricks. 1030 if (padInstructionEncoding(RF, Asm.getEmitter(), RemainingSize)) 1031 FirstChangedFragment = &RF; 1032 1033 // If we have an instruction which hasn't been fully relaxed, we can't 1034 // skip past it and insert bytes before it. Changing its starting 1035 // offset might require a larger negative offset than it can encode. 1036 // We don't need to worry about larger positive offsets as none of the 1037 // possible offsets between this and our align are visible, and the 1038 // ones afterwards aren't changing. 1039 if (!isFullyRelaxed(RF)) 1040 break; 1041 } 1042 Relaxable.clear(); 1043 1044 if (FirstChangedFragment) { 1045 // Make sure the offsets for any fragments in the effected range get 1046 // updated. Note that this (conservatively) invalidates the offsets of 1047 // those following, but this is not required. 1048 Layout.invalidateFragmentsFrom(FirstChangedFragment); 1049 } 1050 1051 // BoundaryAlign explicitly tracks it's size (unlike align) 1052 if (F.getKind() == MCFragment::FT_BoundaryAlign) 1053 cast<MCBoundaryAlignFragment>(F).setSize(RemainingSize); 1054 1055 #ifndef NDEBUG 1056 const uint64_t FinalOffset = Layout.getFragmentOffset(&F); 1057 const uint64_t FinalSize = Asm.computeFragmentSize(Layout, F); 1058 assert(OrigOffset + OrigSize == FinalOffset + FinalSize && 1059 "can't move start of next fragment!"); 1060 assert(FinalSize == RemainingSize && "inconsistent size computation?"); 1061 #endif 1062 1063 // If we're looking at a boundary align, make sure we don't try to pad 1064 // its target instructions for some following directive. Doing so would 1065 // break the alignment of the current boundary align. 1066 if (auto *BF = dyn_cast<MCBoundaryAlignFragment>(&F)) { 1067 const MCFragment *LastFragment = BF->getLastFragment(); 1068 if (!LastFragment) 1069 continue; 1070 while (&*I != LastFragment) 1071 ++I; 1072 } 1073 } 1074 } 1075 1076 // The layout is done. Mark every fragment as valid. 1077 for (unsigned int i = 0, n = Layout.getSectionOrder().size(); i != n; ++i) { 1078 MCSection &Section = *Layout.getSectionOrder()[i]; 1079 Layout.getFragmentOffset(&*Section.getFragmentList().rbegin()); 1080 Asm.computeFragmentSize(Layout, *Section.getFragmentList().rbegin()); 1081 } 1082 } 1083 1084 unsigned X86AsmBackend::getMaximumNopSize() const { 1085 if (STI.hasFeature(X86::Mode16Bit)) 1086 return 4; 1087 if (!STI.hasFeature(X86::FeatureNOPL) && !STI.hasFeature(X86::Mode64Bit)) 1088 return 1; 1089 if (STI.getFeatureBits()[X86::FeatureFast7ByteNOP]) 1090 return 7; 1091 if (STI.getFeatureBits()[X86::FeatureFast15ByteNOP]) 1092 return 15; 1093 if (STI.getFeatureBits()[X86::FeatureFast11ByteNOP]) 1094 return 11; 1095 // FIXME: handle 32-bit mode 1096 // 15-bytes is the longest single NOP instruction, but 10-bytes is 1097 // commonly the longest that can be efficiently decoded. 1098 return 10; 1099 } 1100 1101 /// Write a sequence of optimal nops to the output, covering \p Count 1102 /// bytes. 1103 /// \return - true on success, false on failure 1104 bool X86AsmBackend::writeNopData(raw_ostream &OS, uint64_t Count) const { 1105 static const char Nops32Bit[10][11] = { 1106 // nop 1107 "\x90", 1108 // xchg %ax,%ax 1109 "\x66\x90", 1110 // nopl (%[re]ax) 1111 "\x0f\x1f\x00", 1112 // nopl 0(%[re]ax) 1113 "\x0f\x1f\x40\x00", 1114 // nopl 0(%[re]ax,%[re]ax,1) 1115 "\x0f\x1f\x44\x00\x00", 1116 // nopw 0(%[re]ax,%[re]ax,1) 1117 "\x66\x0f\x1f\x44\x00\x00", 1118 // nopl 0L(%[re]ax) 1119 "\x0f\x1f\x80\x00\x00\x00\x00", 1120 // nopl 0L(%[re]ax,%[re]ax,1) 1121 "\x0f\x1f\x84\x00\x00\x00\x00\x00", 1122 // nopw 0L(%[re]ax,%[re]ax,1) 1123 "\x66\x0f\x1f\x84\x00\x00\x00\x00\x00", 1124 // nopw %cs:0L(%[re]ax,%[re]ax,1) 1125 "\x66\x2e\x0f\x1f\x84\x00\x00\x00\x00\x00", 1126 }; 1127 1128 // 16-bit mode uses different nop patterns than 32-bit. 1129 static const char Nops16Bit[4][11] = { 1130 // nop 1131 "\x90", 1132 // xchg %eax,%eax 1133 "\x66\x90", 1134 // lea 0(%si),%si 1135 "\x8d\x74\x00", 1136 // lea 0w(%si),%si 1137 "\x8d\xb4\x00\x00", 1138 }; 1139 1140 const char(*Nops)[11] = 1141 STI.getFeatureBits()[X86::Mode16Bit] ? Nops16Bit : Nops32Bit; 1142 1143 uint64_t MaxNopLength = (uint64_t)getMaximumNopSize(); 1144 1145 // Emit as many MaxNopLength NOPs as needed, then emit a NOP of the remaining 1146 // length. 1147 do { 1148 const uint8_t ThisNopLength = (uint8_t) std::min(Count, MaxNopLength); 1149 const uint8_t Prefixes = ThisNopLength <= 10 ? 0 : ThisNopLength - 10; 1150 for (uint8_t i = 0; i < Prefixes; i++) 1151 OS << '\x66'; 1152 const uint8_t Rest = ThisNopLength - Prefixes; 1153 if (Rest != 0) 1154 OS.write(Nops[Rest - 1], Rest); 1155 Count -= ThisNopLength; 1156 } while (Count != 0); 1157 1158 return true; 1159 } 1160 1161 /* *** */ 1162 1163 namespace { 1164 1165 class ELFX86AsmBackend : public X86AsmBackend { 1166 public: 1167 uint8_t OSABI; 1168 ELFX86AsmBackend(const Target &T, uint8_t OSABI, const MCSubtargetInfo &STI) 1169 : X86AsmBackend(T, STI), OSABI(OSABI) {} 1170 }; 1171 1172 class ELFX86_32AsmBackend : public ELFX86AsmBackend { 1173 public: 1174 ELFX86_32AsmBackend(const Target &T, uint8_t OSABI, 1175 const MCSubtargetInfo &STI) 1176 : ELFX86AsmBackend(T, OSABI, STI) {} 1177 1178 std::unique_ptr<MCObjectTargetWriter> 1179 createObjectTargetWriter() const override { 1180 return createX86ELFObjectWriter(/*IsELF64*/ false, OSABI, ELF::EM_386); 1181 } 1182 }; 1183 1184 class ELFX86_X32AsmBackend : public ELFX86AsmBackend { 1185 public: 1186 ELFX86_X32AsmBackend(const Target &T, uint8_t OSABI, 1187 const MCSubtargetInfo &STI) 1188 : ELFX86AsmBackend(T, OSABI, STI) {} 1189 1190 std::unique_ptr<MCObjectTargetWriter> 1191 createObjectTargetWriter() const override { 1192 return createX86ELFObjectWriter(/*IsELF64*/ false, OSABI, 1193 ELF::EM_X86_64); 1194 } 1195 }; 1196 1197 class ELFX86_IAMCUAsmBackend : public ELFX86AsmBackend { 1198 public: 1199 ELFX86_IAMCUAsmBackend(const Target &T, uint8_t OSABI, 1200 const MCSubtargetInfo &STI) 1201 : ELFX86AsmBackend(T, OSABI, STI) {} 1202 1203 std::unique_ptr<MCObjectTargetWriter> 1204 createObjectTargetWriter() const override { 1205 return createX86ELFObjectWriter(/*IsELF64*/ false, OSABI, 1206 ELF::EM_IAMCU); 1207 } 1208 }; 1209 1210 class ELFX86_64AsmBackend : public ELFX86AsmBackend { 1211 public: 1212 ELFX86_64AsmBackend(const Target &T, uint8_t OSABI, 1213 const MCSubtargetInfo &STI) 1214 : ELFX86AsmBackend(T, OSABI, STI) {} 1215 1216 std::unique_ptr<MCObjectTargetWriter> 1217 createObjectTargetWriter() const override { 1218 return createX86ELFObjectWriter(/*IsELF64*/ true, OSABI, ELF::EM_X86_64); 1219 } 1220 }; 1221 1222 class WindowsX86AsmBackend : public X86AsmBackend { 1223 bool Is64Bit; 1224 1225 public: 1226 WindowsX86AsmBackend(const Target &T, bool is64Bit, 1227 const MCSubtargetInfo &STI) 1228 : X86AsmBackend(T, STI) 1229 , Is64Bit(is64Bit) { 1230 } 1231 1232 Optional<MCFixupKind> getFixupKind(StringRef Name) const override { 1233 return StringSwitch<Optional<MCFixupKind>>(Name) 1234 .Case("dir32", FK_Data_4) 1235 .Case("secrel32", FK_SecRel_4) 1236 .Case("secidx", FK_SecRel_2) 1237 .Default(MCAsmBackend::getFixupKind(Name)); 1238 } 1239 1240 std::unique_ptr<MCObjectTargetWriter> 1241 createObjectTargetWriter() const override { 1242 return createX86WinCOFFObjectWriter(Is64Bit); 1243 } 1244 }; 1245 1246 namespace CU { 1247 1248 /// Compact unwind encoding values. 1249 enum CompactUnwindEncodings { 1250 /// [RE]BP based frame where [RE]BP is pused on the stack immediately after 1251 /// the return address, then [RE]SP is moved to [RE]BP. 1252 UNWIND_MODE_BP_FRAME = 0x01000000, 1253 1254 /// A frameless function with a small constant stack size. 1255 UNWIND_MODE_STACK_IMMD = 0x02000000, 1256 1257 /// A frameless function with a large constant stack size. 1258 UNWIND_MODE_STACK_IND = 0x03000000, 1259 1260 /// No compact unwind encoding is available. 1261 UNWIND_MODE_DWARF = 0x04000000, 1262 1263 /// Mask for encoding the frame registers. 1264 UNWIND_BP_FRAME_REGISTERS = 0x00007FFF, 1265 1266 /// Mask for encoding the frameless registers. 1267 UNWIND_FRAMELESS_STACK_REG_PERMUTATION = 0x000003FF 1268 }; 1269 1270 } // namespace CU 1271 1272 class DarwinX86AsmBackend : public X86AsmBackend { 1273 const MCRegisterInfo &MRI; 1274 1275 /// Number of registers that can be saved in a compact unwind encoding. 1276 enum { CU_NUM_SAVED_REGS = 6 }; 1277 1278 mutable unsigned SavedRegs[CU_NUM_SAVED_REGS]; 1279 Triple TT; 1280 bool Is64Bit; 1281 1282 unsigned OffsetSize; ///< Offset of a "push" instruction. 1283 unsigned MoveInstrSize; ///< Size of a "move" instruction. 1284 unsigned StackDivide; ///< Amount to adjust stack size by. 1285 protected: 1286 /// Size of a "push" instruction for the given register. 1287 unsigned PushInstrSize(unsigned Reg) const { 1288 switch (Reg) { 1289 case X86::EBX: 1290 case X86::ECX: 1291 case X86::EDX: 1292 case X86::EDI: 1293 case X86::ESI: 1294 case X86::EBP: 1295 case X86::RBX: 1296 case X86::RBP: 1297 return 1; 1298 case X86::R12: 1299 case X86::R13: 1300 case X86::R14: 1301 case X86::R15: 1302 return 2; 1303 } 1304 return 1; 1305 } 1306 1307 private: 1308 /// Get the compact unwind number for a given register. The number 1309 /// corresponds to the enum lists in compact_unwind_encoding.h. 1310 int getCompactUnwindRegNum(unsigned Reg) const { 1311 static const MCPhysReg CU32BitRegs[7] = { 1312 X86::EBX, X86::ECX, X86::EDX, X86::EDI, X86::ESI, X86::EBP, 0 1313 }; 1314 static const MCPhysReg CU64BitRegs[] = { 1315 X86::RBX, X86::R12, X86::R13, X86::R14, X86::R15, X86::RBP, 0 1316 }; 1317 const MCPhysReg *CURegs = Is64Bit ? CU64BitRegs : CU32BitRegs; 1318 for (int Idx = 1; *CURegs; ++CURegs, ++Idx) 1319 if (*CURegs == Reg) 1320 return Idx; 1321 1322 return -1; 1323 } 1324 1325 /// Return the registers encoded for a compact encoding with a frame 1326 /// pointer. 1327 uint32_t encodeCompactUnwindRegistersWithFrame() const { 1328 // Encode the registers in the order they were saved --- 3-bits per 1329 // register. The list of saved registers is assumed to be in reverse 1330 // order. The registers are numbered from 1 to CU_NUM_SAVED_REGS. 1331 uint32_t RegEnc = 0; 1332 for (int i = 0, Idx = 0; i != CU_NUM_SAVED_REGS; ++i) { 1333 unsigned Reg = SavedRegs[i]; 1334 if (Reg == 0) break; 1335 1336 int CURegNum = getCompactUnwindRegNum(Reg); 1337 if (CURegNum == -1) return ~0U; 1338 1339 // Encode the 3-bit register number in order, skipping over 3-bits for 1340 // each register. 1341 RegEnc |= (CURegNum & 0x7) << (Idx++ * 3); 1342 } 1343 1344 assert((RegEnc & 0x3FFFF) == RegEnc && 1345 "Invalid compact register encoding!"); 1346 return RegEnc; 1347 } 1348 1349 /// Create the permutation encoding used with frameless stacks. It is 1350 /// passed the number of registers to be saved and an array of the registers 1351 /// saved. 1352 uint32_t encodeCompactUnwindRegistersWithoutFrame(unsigned RegCount) const { 1353 // The saved registers are numbered from 1 to 6. In order to encode the 1354 // order in which they were saved, we re-number them according to their 1355 // place in the register order. The re-numbering is relative to the last 1356 // re-numbered register. E.g., if we have registers {6, 2, 4, 5} saved in 1357 // that order: 1358 // 1359 // Orig Re-Num 1360 // ---- ------ 1361 // 6 6 1362 // 2 2 1363 // 4 3 1364 // 5 3 1365 // 1366 for (unsigned i = 0; i < RegCount; ++i) { 1367 int CUReg = getCompactUnwindRegNum(SavedRegs[i]); 1368 if (CUReg == -1) return ~0U; 1369 SavedRegs[i] = CUReg; 1370 } 1371 1372 // Reverse the list. 1373 std::reverse(&SavedRegs[0], &SavedRegs[CU_NUM_SAVED_REGS]); 1374 1375 uint32_t RenumRegs[CU_NUM_SAVED_REGS]; 1376 for (unsigned i = CU_NUM_SAVED_REGS - RegCount; i < CU_NUM_SAVED_REGS; ++i){ 1377 unsigned Countless = 0; 1378 for (unsigned j = CU_NUM_SAVED_REGS - RegCount; j < i; ++j) 1379 if (SavedRegs[j] < SavedRegs[i]) 1380 ++Countless; 1381 1382 RenumRegs[i] = SavedRegs[i] - Countless - 1; 1383 } 1384 1385 // Take the renumbered values and encode them into a 10-bit number. 1386 uint32_t permutationEncoding = 0; 1387 switch (RegCount) { 1388 case 6: 1389 permutationEncoding |= 120 * RenumRegs[0] + 24 * RenumRegs[1] 1390 + 6 * RenumRegs[2] + 2 * RenumRegs[3] 1391 + RenumRegs[4]; 1392 break; 1393 case 5: 1394 permutationEncoding |= 120 * RenumRegs[1] + 24 * RenumRegs[2] 1395 + 6 * RenumRegs[3] + 2 * RenumRegs[4] 1396 + RenumRegs[5]; 1397 break; 1398 case 4: 1399 permutationEncoding |= 60 * RenumRegs[2] + 12 * RenumRegs[3] 1400 + 3 * RenumRegs[4] + RenumRegs[5]; 1401 break; 1402 case 3: 1403 permutationEncoding |= 20 * RenumRegs[3] + 4 * RenumRegs[4] 1404 + RenumRegs[5]; 1405 break; 1406 case 2: 1407 permutationEncoding |= 5 * RenumRegs[4] + RenumRegs[5]; 1408 break; 1409 case 1: 1410 permutationEncoding |= RenumRegs[5]; 1411 break; 1412 } 1413 1414 assert((permutationEncoding & 0x3FF) == permutationEncoding && 1415 "Invalid compact register encoding!"); 1416 return permutationEncoding; 1417 } 1418 1419 public: 1420 DarwinX86AsmBackend(const Target &T, const MCRegisterInfo &MRI, 1421 const MCSubtargetInfo &STI) 1422 : X86AsmBackend(T, STI), MRI(MRI), TT(STI.getTargetTriple()), 1423 Is64Bit(TT.isArch64Bit()) { 1424 memset(SavedRegs, 0, sizeof(SavedRegs)); 1425 OffsetSize = Is64Bit ? 8 : 4; 1426 MoveInstrSize = Is64Bit ? 3 : 2; 1427 StackDivide = Is64Bit ? 8 : 4; 1428 } 1429 1430 std::unique_ptr<MCObjectTargetWriter> 1431 createObjectTargetWriter() const override { 1432 uint32_t CPUType = cantFail(MachO::getCPUType(TT)); 1433 uint32_t CPUSubType = cantFail(MachO::getCPUSubType(TT)); 1434 return createX86MachObjectWriter(Is64Bit, CPUType, CPUSubType); 1435 } 1436 1437 /// Implementation of algorithm to generate the compact unwind encoding 1438 /// for the CFI instructions. 1439 uint32_t 1440 generateCompactUnwindEncoding(ArrayRef<MCCFIInstruction> Instrs) const override { 1441 if (Instrs.empty()) return 0; 1442 1443 // Reset the saved registers. 1444 unsigned SavedRegIdx = 0; 1445 memset(SavedRegs, 0, sizeof(SavedRegs)); 1446 1447 bool HasFP = false; 1448 1449 // Encode that we are using EBP/RBP as the frame pointer. 1450 uint32_t CompactUnwindEncoding = 0; 1451 1452 unsigned SubtractInstrIdx = Is64Bit ? 3 : 2; 1453 unsigned InstrOffset = 0; 1454 unsigned StackAdjust = 0; 1455 unsigned StackSize = 0; 1456 unsigned NumDefCFAOffsets = 0; 1457 int MinAbsOffset = std::numeric_limits<int>::max(); 1458 1459 for (unsigned i = 0, e = Instrs.size(); i != e; ++i) { 1460 const MCCFIInstruction &Inst = Instrs[i]; 1461 1462 switch (Inst.getOperation()) { 1463 default: 1464 // Any other CFI directives indicate a frame that we aren't prepared 1465 // to represent via compact unwind, so just bail out. 1466 return 0; 1467 case MCCFIInstruction::OpDefCfaRegister: { 1468 // Defines a frame pointer. E.g. 1469 // 1470 // movq %rsp, %rbp 1471 // L0: 1472 // .cfi_def_cfa_register %rbp 1473 // 1474 HasFP = true; 1475 1476 // If the frame pointer is other than esp/rsp, we do not have a way to 1477 // generate a compact unwinding representation, so bail out. 1478 if (*MRI.getLLVMRegNum(Inst.getRegister(), true) != 1479 (Is64Bit ? X86::RBP : X86::EBP)) 1480 return 0; 1481 1482 // Reset the counts. 1483 memset(SavedRegs, 0, sizeof(SavedRegs)); 1484 StackAdjust = 0; 1485 SavedRegIdx = 0; 1486 MinAbsOffset = std::numeric_limits<int>::max(); 1487 InstrOffset += MoveInstrSize; 1488 break; 1489 } 1490 case MCCFIInstruction::OpDefCfaOffset: { 1491 // Defines a new offset for the CFA. E.g. 1492 // 1493 // With frame: 1494 // 1495 // pushq %rbp 1496 // L0: 1497 // .cfi_def_cfa_offset 16 1498 // 1499 // Without frame: 1500 // 1501 // subq $72, %rsp 1502 // L0: 1503 // .cfi_def_cfa_offset 80 1504 // 1505 StackSize = Inst.getOffset() / StackDivide; 1506 ++NumDefCFAOffsets; 1507 break; 1508 } 1509 case MCCFIInstruction::OpOffset: { 1510 // Defines a "push" of a callee-saved register. E.g. 1511 // 1512 // pushq %r15 1513 // pushq %r14 1514 // pushq %rbx 1515 // L0: 1516 // subq $120, %rsp 1517 // L1: 1518 // .cfi_offset %rbx, -40 1519 // .cfi_offset %r14, -32 1520 // .cfi_offset %r15, -24 1521 // 1522 if (SavedRegIdx == CU_NUM_SAVED_REGS) 1523 // If there are too many saved registers, we cannot use a compact 1524 // unwind encoding. 1525 return CU::UNWIND_MODE_DWARF; 1526 1527 unsigned Reg = *MRI.getLLVMRegNum(Inst.getRegister(), true); 1528 SavedRegs[SavedRegIdx++] = Reg; 1529 StackAdjust += OffsetSize; 1530 MinAbsOffset = std::min(MinAbsOffset, abs(Inst.getOffset())); 1531 InstrOffset += PushInstrSize(Reg); 1532 break; 1533 } 1534 } 1535 } 1536 1537 StackAdjust /= StackDivide; 1538 1539 if (HasFP) { 1540 if ((StackAdjust & 0xFF) != StackAdjust) 1541 // Offset was too big for a compact unwind encoding. 1542 return CU::UNWIND_MODE_DWARF; 1543 1544 // We don't attempt to track a real StackAdjust, so if the saved registers 1545 // aren't adjacent to rbp we can't cope. 1546 if (SavedRegIdx != 0 && MinAbsOffset != 3 * (int)OffsetSize) 1547 return CU::UNWIND_MODE_DWARF; 1548 1549 // Get the encoding of the saved registers when we have a frame pointer. 1550 uint32_t RegEnc = encodeCompactUnwindRegistersWithFrame(); 1551 if (RegEnc == ~0U) return CU::UNWIND_MODE_DWARF; 1552 1553 CompactUnwindEncoding |= CU::UNWIND_MODE_BP_FRAME; 1554 CompactUnwindEncoding |= (StackAdjust & 0xFF) << 16; 1555 CompactUnwindEncoding |= RegEnc & CU::UNWIND_BP_FRAME_REGISTERS; 1556 } else { 1557 SubtractInstrIdx += InstrOffset; 1558 ++StackAdjust; 1559 1560 if ((StackSize & 0xFF) == StackSize) { 1561 // Frameless stack with a small stack size. 1562 CompactUnwindEncoding |= CU::UNWIND_MODE_STACK_IMMD; 1563 1564 // Encode the stack size. 1565 CompactUnwindEncoding |= (StackSize & 0xFF) << 16; 1566 } else { 1567 if ((StackAdjust & 0x7) != StackAdjust) 1568 // The extra stack adjustments are too big for us to handle. 1569 return CU::UNWIND_MODE_DWARF; 1570 1571 // Frameless stack with an offset too large for us to encode compactly. 1572 CompactUnwindEncoding |= CU::UNWIND_MODE_STACK_IND; 1573 1574 // Encode the offset to the nnnnnn value in the 'subl $nnnnnn, ESP' 1575 // instruction. 1576 CompactUnwindEncoding |= (SubtractInstrIdx & 0xFF) << 16; 1577 1578 // Encode any extra stack adjustments (done via push instructions). 1579 CompactUnwindEncoding |= (StackAdjust & 0x7) << 13; 1580 } 1581 1582 // Encode the number of registers saved. (Reverse the list first.) 1583 std::reverse(&SavedRegs[0], &SavedRegs[SavedRegIdx]); 1584 CompactUnwindEncoding |= (SavedRegIdx & 0x7) << 10; 1585 1586 // Get the encoding of the saved registers when we don't have a frame 1587 // pointer. 1588 uint32_t RegEnc = encodeCompactUnwindRegistersWithoutFrame(SavedRegIdx); 1589 if (RegEnc == ~0U) return CU::UNWIND_MODE_DWARF; 1590 1591 // Encode the register encoding. 1592 CompactUnwindEncoding |= 1593 RegEnc & CU::UNWIND_FRAMELESS_STACK_REG_PERMUTATION; 1594 } 1595 1596 return CompactUnwindEncoding; 1597 } 1598 }; 1599 1600 } // end anonymous namespace 1601 1602 MCAsmBackend *llvm::createX86_32AsmBackend(const Target &T, 1603 const MCSubtargetInfo &STI, 1604 const MCRegisterInfo &MRI, 1605 const MCTargetOptions &Options) { 1606 const Triple &TheTriple = STI.getTargetTriple(); 1607 if (TheTriple.isOSBinFormatMachO()) 1608 return new DarwinX86AsmBackend(T, MRI, STI); 1609 1610 if (TheTriple.isOSWindows() && TheTriple.isOSBinFormatCOFF()) 1611 return new WindowsX86AsmBackend(T, false, STI); 1612 1613 uint8_t OSABI = MCELFObjectTargetWriter::getOSABI(TheTriple.getOS()); 1614 1615 if (TheTriple.isOSIAMCU()) 1616 return new ELFX86_IAMCUAsmBackend(T, OSABI, STI); 1617 1618 return new ELFX86_32AsmBackend(T, OSABI, STI); 1619 } 1620 1621 MCAsmBackend *llvm::createX86_64AsmBackend(const Target &T, 1622 const MCSubtargetInfo &STI, 1623 const MCRegisterInfo &MRI, 1624 const MCTargetOptions &Options) { 1625 const Triple &TheTriple = STI.getTargetTriple(); 1626 if (TheTriple.isOSBinFormatMachO()) 1627 return new DarwinX86AsmBackend(T, MRI, STI); 1628 1629 if (TheTriple.isOSWindows() && TheTriple.isOSBinFormatCOFF()) 1630 return new WindowsX86AsmBackend(T, true, STI); 1631 1632 uint8_t OSABI = MCELFObjectTargetWriter::getOSABI(TheTriple.getOS()); 1633 1634 if (TheTriple.isX32()) 1635 return new ELFX86_X32AsmBackend(T, OSABI, STI); 1636 return new ELFX86_64AsmBackend(T, OSABI, STI); 1637 } 1638